The recent hype about artificial intelligence has sparked renewed interest in applying the successful deep learning (DL) methods for image recognition, speech recognition, robotics, strategic games and other application areas to the field of meteorology. There is some evidence that better weather forecasts can be produced by introducing big data mining and neural networks into the weather prediction workflow. Here, we discuss the question of whether it is possible to completely replace the current numerical weather models and data assimilation systems with DL approaches. This discussion entails a review of state-of-the-art machine learning concepts and their applicability to weather data with its pertinent statistical properties. We think that it is not inconceivable that numerical weather models may one day become obsolete, but a number of fundamental breakthroughs are needed before this goal comes into reach. This article is part of the theme issue ‘Machine learning for weather and climate modelling’.

In this paper, we present the implementation and evaluation of the aerosol microphysics module SALSA2.0 in the framework of the aerosol–chemistry–climate model ECHAM-HAMMOZ. It is an alternative microphysics module to the default modal microphysics scheme M7 in ECHAM-HAMMOZ. The SALSA2.0 implementation within ECHAM-HAMMOZ is evaluated against observations of aerosol optical properties, aerosol mass, and size distributions, comparing also to the skill of the M7 implementation. The largest differences between the implementation of SALSA2.0 and M7 are in the methods used for calculating microphysical processes, i.e., nucleation, condensation, coagulation, and hydration. These differences in the microphysics are reflected in the results so that the largest differences between SALSA2.0 and M7 are evident over regions where the aerosol size distribution is heavily modified by the microphysical processing of aerosol particles. Such regions are, for example, highly polluted regions and regions strongly affected by biomass burning. In addition, in a simulation of the 1991 Mt. Pinatubo eruption in which a stratospheric sulfate plume was formed, the global burden and the effective radii of the stratospheric aerosol are very different in SALSA2.0 and M7. While SALSA2.0 was able to reproduce the observed time evolution of the global burden of sulfate and the effective radii of stratospheric aerosol, M7 strongly overestimates the removal of coarse stratospheric particles and thus underestimates the effective radius of stratospheric aerosol. As the mode widths of M7 have been optimized for the troposphere and were not designed to represent stratospheric aerosol, the ability of M7 to simulate the volcano plume was improved by modifying the mode widths, decreasing the standard deviations of the accumulation and coarse modes from 1.59 and 2.0, respectively, to 1.2 similar to what was observed after the Mt. Pinatubo eruption. Overall, SALSA2.0 shows promise in improving the aerosol description of ECHAM-HAMMOZ and can be further improved by implementing methods for aerosol processes that are more suitable for the sectional method, e.g., size-dependent emissions for aerosol species and size-resolved wet deposition.

The chemistry–climate model ECHAM-HAMMOZ contains a detailed representation of tropospheric and stratospheric reactive chemistry and state-of-the-art parameterizations of aerosols using either a modal scheme (M7) or a bin scheme (SALSA). This article describes and evaluates the model version ECHAM6.3-HAM2.3-MOZ1.0 with a focus on the tropospheric gas-phase chemistry. A 10-year model simulation was performed to test the stability of the model and provide data for its evaluation. The comparison to observations concentrates on the year 2008 and includes total column observations of ozone and CO from IASI and OMI, Aura MLS observations of temperature, HNO₃, ClO, and O₃ for the evaluation of polar stratospheric processes, an ozonesonde climatology, surface ozone observations from the TOAR database, and surface CO data from the Global Atmosphere Watch network. Global budgets of ozone, OH, NO_x, aerosols, clouds, and radiation are analyzed and compared to the literature. ECHAM-HAMMOZ performs well in many aspects. However, in the base simulation, lightning NO_x emissions are very low, and the impact of the heterogeneous reaction of HNO₃ on dust and sea salt aerosol is too strong. Sensitivity simulations with increased lightning NO_x or modified heterogeneous chemistry deteriorate the comparison with observations and yield excessively large ozone budget terms and too much OH. We hypothesize that this is an impact of potential issues with tropical convection in the ECHAM model.

Numerical weather prediction (NWP) models solve a system of partial differential equations based on physical laws to forecast the future state of the atmosphere. These models are deployed operationally, but they are computationally very expensive. Recently, the potential of deep neural networks to generate bespoke weather forecasts has been explored in a couple of scientific studies inspired by the success of video frame prediction models in computer vision. In this study, a simple recurrent neural network with convolutional filters, called ConvLSTM, and an advanced generative network, the Stochastic Adversarial Video Prediction (SAVP) model, are applied to create hourly forecasts of the 2 m temperature for the next 12 h over Europe. We make use of 13 years of data from the ERA5 reanalysis, of which 11 years are utilized for training and 1 year each is used for validating and testing. We choose the 2 m temperature, total cloud cover, and the 850 hPa temperature as predictors and show that both models attain predictive skill by outperforming persistence forecasts. SAVP is superior to ConvLSTM in terms of several evaluation metrics, confirming previous results from computer vision that larger, more complex networks are better suited to learn complex features and to generate better predictions. The 12 h forecasts of SAVP attain a mean squared error (MSE) of about 2.3 K², an anomaly correlation coefficient (ACC) larger than 0.85, a structural similarity index (SSIM) of around 0.72, and a gradient ratio (r_G) of about 0.82. The ConvLSTM yields a higher MSE (3.6 K²), a smaller ACC (0.80) and SSIM (0.65), and a slightly larger r_G (0.84). The superior performance of SAVP in terms of MSE, ACC, and SSIM can be largely attributed to the generator. A sensitivity study shows that a larger weight of the generative adversarial network (GAN) component in the SAVP loss leads to even better preservation of spatial variability at the cost of a somewhat increased MSE (2.5 K²). Including the 850 hPa temperature as an additional predictor enhances the forecast quality, and the model also benefits from a larger spatial domain. By contrast, adding the total cloud cover as predictor or reducing the amount of training data to 8 years has only small effects. Although the temperature forecasts obtained in this way are still less powerful than contemporary NWP models, this study demonstrates that sophisticated deep neural networks may achieve considerable forecast quality beyond the nowcasting range in a purely data-driven way.

The impact of six heterogeneous gas–aerosol uptake reactions on tropospheric ozone and nitrogen species was studied using two chemical transport models, the Meteorological Synthesizing Centre-West of the European Monitoring and Evaluation Programme (EMEP MSC-W) and the European Centre Hamburg general circulation model combined with versions of the Hamburg Aerosol Model and Model for Ozone and Related chemical Tracers (ECHAM-HAMMOZ). Species undergoing heterogeneous reactions in both models include N₂O₅, NO₃, NO₂, O₃, HNO₃, and HO₂. Since heterogeneous reactions take place at the aerosol surface area, the modelled surface area density (S_a) of both models was compared to a satellite product retrieving the surface area. This comparison shows a good agreement in global pattern and especially the capability of both models to capture the extreme aerosol loadings in east Asia. The impact of the heterogeneous reactions was evaluated by the simulation of a reference run containing all heterogeneous reactions and several sensitivity runs. One reaction was turned off in each sensitivity run to compare it with the reference run. The analysis of the sensitivity runs confirms that the globally most important heterogeneous reaction is the one of N₂O₅. Nevertheless, NO₂, HNO₃, and HO₂ heterogeneous reactions gain relevance particularly in east Asia due to the presence of high NO_x concentrations and high S_a in the same region. The heterogeneous reaction of O₃ itself on dust is of minor relevance compared to the other heterogeneous reactions. The impacts of the N₂O₅ reactions show strong seasonal variations, with the biggest impacts on O₃ in springtime when photochemical reactions are active and N₂O₅ levels still high. Evaluation of the models with northern hemispheric ozone surface observations yields a better agreement of the models with observations in terms of concentration levels, variability, and temporal correlations at most sites when the heterogeneous reactions are incorporated. Our results are loosely consistent with results from earlier studies, although the magnitude of changes induced by N₂O₅ reaction is at the low end of estimates, which seems to fit a trend, whereby the more recent the study the lower the impacts of these reactions.

Tropospheric ozone is a toxic greenhouse gas with a highly variable spatial distribution which is challenging to map on a global scale. Here, we present a data-driven ozone-mapping workflow generating a transparent and reliable product. We map the global distribution of tropospheric ozone from sparse, irregularly placed measurement stations to a high-resolution regular grid using machine learning methods. The produced map contains the average tropospheric ozone concentration of the years 2010–2014 with a resolution of 0.1^∘ × 0.1^∘. The machine learning model is trained on AQ-Bench (“air quality benchmark dataset”), a pre-compiled benchmark dataset consisting of multi-year ground-based ozone measurements combined with an abundance of high-resolution geospatial data. Going beyond standard mapping methods, this work focuses on two key aspects to increase the integrity of the produced map. Using explainable machine learning methods, we ensure that the trained machine learning model is consistent with commonly accepted knowledge about tropospheric ozone. To assess the impact of data and model uncertainties on our ozone map, we show that the machine learning model is robust against typical fluctuations in ozone values and geospatial data. By inspecting the input features, we ensure that the model is only applied in regions where it is reliable. We provide a rationale for the tools we use to conduct a thorough global analysis. The methods presented here can thus be easily transferred to other mapping applications to ensure the transparency and reliability of the maps produced.